What is Ministral 3 (3B, 8B and 14B)?
Ministral 3 is Mistral AI's family of small models, released on 2 December 2025 alongside Mistral Large 3. It comes in three dense sizes, 3B, 8B and 14B parameters, and each size has a base, an instruct and a reasoning version. All of them read images as well as text, support a 256K context (128K on the 3B via OpenRouter) and are licensed under Apache 2.0.
Mistral built them for local and edge use. The 14B instruct model fits in 24 GB of video memory in FP8, and NVIDIA has optimised the family for DGX Spark, RTX PCs and laptops, and Jetson devices. The weights are on Hugging Face, and the models are on Mistral's API and cloud platforms including Amazon Bedrock, Azure Foundry, OpenRouter and Together AI.
On Mistral's API the 14B costs $0.20 per million tokens, the 8B $0.15 and the 3B $0.10, with the same rate for input and output. Mistral says the 14B performs comparably to its older, larger Mistral Small 3.2 24B.
How much does Ministral 3 (3B, 8B and 14B) cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0 |
| Ministral 3 14B | $0.20 per 1M tokens | Input and output |
| Ministral 3 8B | $0.15 per 1M tokens | |
| Ministral 3 3B | $0.10 per 1M tokens |
Open weights under Apache 2.0. Mistral API: 14B $0.20, 8B $0.15, 3B $0.10 per 1M tokens, the same for input and output; batch half price.
Is Ministral 3 (3B, 8B and 14B) worth it?
Ministral 3's strength is completeness. Few families offer three sizes, each in base, instruct and reasoning versions, all reading images, all under Apache 2.0 with no conditions. For a team that wants to fine-tune a base model, ship an instruct model and fall back to a reasoning model for hard queries, from one vendor and licence, it covers everything.
On Mistral's own tables the reasoning versions are strong for their size. The 14B reasoning model scores 85.0% on AIME 2025 and 71.2% on GPQA Diamond, and the 3B beats Qwen3-VL-4B-Thinking on AIME. The instruct 14B leads Qwen3 14B and Gemma 3 12B on Arena Hard and WildBench. But Qwen's vision models win several of the head-to-heads Mistral prints: Qwen3-VL-8B-Instruct beats the 8B on Arena Hard and maths, and Qwen3-VL-4B-Instruct beats the 3B on most rows.
The local story is the practical one. The 14B instruct runs in 24 GB in FP8, so a single consumer GPU or a recent laptop with enough memory will do, and the 3B fits on phones and Jetson boards. Mistral says the instruct models often produce far fewer tokens than rivals for the same answer, which matters for latency on small hardware.
It is ten months old, and newer small models have arrived since, including Gemma 4. It remains a safe default for European teams wanting a permissive, multimodal small model.
Where it earns its place
A developer building an offline assistant for a laptop app
Run a capable model locally with image input
Use Ministral 3 14B Instruct in FP8, which fits in 24 GB, or quantise further. NVIDIA's RTX and DGX Spark optimisations cover the common hardware.
A robotics or IoT team
Put a vision-language model on an embedded board
Ministral 3 3B runs on NVIDIA Jetson devices and reads camera frames directly. Switch to the 3B reasoning version for multi-step planning.
A startup classifying and extracting from documents at volume
Keep per-request costs near zero
On Mistral's API, the 8B at $0.15 per million tokens handles a million 1,500-token extraction calls for about $225, or half on batch.
Before you commit
- Setup
- Download from Hugging Face (instruct models in FP8, base and reasoning in BF16), or call ministral-3b-2512, ministral-8b-2512 or ministral-14b-2512 on Mistral's API.
- Integrations
- vLLM, SGLang, TensorRT-LLM, llama.cpp and Ollama builds; Amazon Bedrock, Azure Foundry, IBM watsonx, OpenRouter, Fireworks, Together AI and Modal. Function calling and structured outputs.
- Data and privacy
- Apache 2.0 with no usage restrictions when self-hosted. On Mistral's API, free-tier inputs and outputs may be used for training unless you opt out; pay-as-you-go users can opt out.
- Cost at scale
- $0.10 to $0.20 per million tokens on Mistral's API, with batch at half price. Self-hosted, the 14B in FP8 needs one 24 GB GPU; the 3B runs on phones and Jetson boards.
- Limits
- 256K context on the 8B and 14B (128K for the 3B on OpenRouter); text and image input only. Reasoning versions use more tokens than the instruct ones.
What's good
- Base, instruct and reasoning versions at 3B, 8B and 14B
- Image input in every model; 256K context
- Apache 2.0 with no conditions
- 14B instruct fits in 24 GB of video memory in FP8
What's not
- Qwen3-VL models beat it on several of Mistral's own comparisons
- Released December 2025; newer small models have arrived since
- Mistral's API free tier may train on inputs unless you opt out
- No audio input
Best for
Teams that need a small, permissively licensed model with vision for laptops, edge devices or cheap hosted inference.
Ministral 3 (3B, 8B and 14B) alternatives
Ministral 3 (3B, 8B and 14B): common questions
- Is Ministral 3 (3B, 8B and 14B) worth it?
- A complete small-model family under a clean licence, with reasoning and vision in every size; Qwen still edges it on several of Mistral's own comparisons. We rate it 4.1 out of 5.
- How much does Ministral 3 (3B, 8B and 14B) cost?
- Open weights under Apache 2.0. Mistral API: 14B $0.20, 8B $0.15, 3B $0.10 per 1M tokens, the same for input and output; batch half price.
- What are the downsides of Ministral 3 (3B, 8B and 14B)?
- Qwen3-VL models beat it on several of Mistral's own comparisons; released December 2025; newer small models have arrived since; mistral's API free tier may train on inputs unless you opt out; no audio input.
- Who is Ministral 3 (3B, 8B and 14B) best for?
- Teams that need a small, permissively licensed model with vision for laptops, edge devices or cheap hosted inference.
- How much does Ministral 3 (3B, 8B and 14B) cost for a team or in production?
- $0.10 to $0.20 per million tokens on Mistral's API, with batch at half price. Self-hosted, the 14B in FP8 needs one 24 GB GPU; the 3B runs on phones and Jetson boards.
- Is Ministral 3 (3B, 8B and 14B) safe for company data?
- Apache 2.0 with no usage restrictions when self-hosted. On Mistral's API, free-tier inputs and outputs may be used for training unless you opt out; pay-as-you-go users can opt out.
- What are the best Ministral 3 (3B, 8B and 14B) alternatives?
- The closest alternatives we have reviewed are Gemma 4, Phi-4-multimodal, MiniCPM5-2B and SmolLM3-3B.
Answers checked Oct 25, 2026.
Make Ministral 3 (3B, 8B and 14B)? Show this rating on your site
